Subagents
A subagent is a worker your main session spawns for a focused task. It has its own context window, does its job, and reports a summary back. Subagents only talk to the parent, not to each other. The docs recommend them for “focused tasks where only the result matters”, and you can give one an isolated copy of the repo with isolation: worktree.
Agent teams
Agent teams are experimental and off by default; you enable them with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. One session is the lead; teammates each have their own context, claim items from a shared task list, and message each other through a mailbox. Teams and tasks are stored locally under ~/.claude/teams/, a session has exactly one team, and the docs say to avoid two teammates editing the same file because it “leads to overwrites”. They also warn that teams use “significantly more tokens than a single session”.
What both have in common
Both are one person’s Claude, multiplied. The lead is your session, the teammates are on your machine, the task list is in your home folder. That is exactly right for fanning one job out. It says nothing about the Claude your colleague started on their laptop ten minutes ago, on the same repo.
A co-work
A Chain co-work is the layer for that case. The repo gets a shareable code. Each person runs their own agent, on their own Mac, with their own keys. All of those agents, whether Claude Code, Codex or Cursor, read and write one board of tasks, claims and log lines. A pre-tool hook refuses edits to paths another agent holds. When an agent is done, Chain tests, merges, pushes and releases its claims; a real conflict pauses the merge for a person. Agent teams still work inside a co-work: a lead and its teammates on your Mac are just more agents that respect the same claims.
Side by side
| Chain | Agent teams | |
|---|---|---|
| People | Everyone in the co-work | One person (the lead’s session) |
| Machines | Every member’s Mac | One machine |
| Agents | Claude Code, Codex CLI, Cursor, OpenClaw, any MCP agent | Claude Code only |
| Shared state | Board of tasks, claims and log, synced through the relay | Local task list and mailboxes under ~/.claude |
| Same-file protection | Path claims enforced by a hook | Advice to split files between teammates |
| Merging | Automatic when tests pass; pauses on conflict | Not covered; you merge yourself |
| Status | Private beta, macOS | Experimental, opt-in with an env var |
Agent team and subagent details are from the Claude Code documentation as of October 2026.
Which one, when
- One task, one person: a single session, maybe with subagents for research or review.
- One big job, one person, several angles at once: agent teams, with each teammate owning different files.
- More than one person running agents on the repo: a co-work, with whatever each person runs inside it. The Chain for Claude Code page shows how it is wired in.